Research And Application Of Spindle Position Detection Based On Deep Learning

With the development of convolutional neural networks and the rapid development of deep learning technology,computer vision technology has been integrated into the pace of intelligence in various industries.Object position detection is one of the core research directions of computer vision.The application of object position detection technology in industrial production can effectively improve the efficiency of production and processing,and help or even replace humans in complex environments.In this paper,object detection is applied to the spindle production line to reduce the manpower requirements for working in this high humidity and hot environment,and effectively improve production efficiency.In the process of spindle production and processing,in order to achieve different goals and different production needs,it is necessary to manually sort the shaped or semi-formed spindles,and then transport and place them in the corresponding subsequent positions,which may even require artificial Arrange the spindles in a regular position to meet subsequent processing requirements.This undoubtedly has a huge impact on the production efficiency of the spindle.This article uses deep learning to detect the position of the spindle to provide the robot with the specific three-dimensional position information of the spindle,to assist it in sorting and transportation,so as to improve the production efficiency of the spindle and reduce the waste of human resources.Aiming at the actual situation and real-time requirements of disorderly stacked spindles,this paper uses two-dimensional image detection combined with the corresponding spindle depth image to detect the three-dimensional spatial information of the spindle.Firstly,the key points,positions and categories of the spindles in the color image are detected,and the candidate spindles are selected according to the detection results,and then the candidate spindles are obtained from the depth map corresponding to the color image according to the detected two-dimensional spatial information.Depth information,to calculate the rotation and translation of the candidate spindle in depth space,and finally the two-dimensional spatial information of the candidate spindle and the depth space information are fused by the focal length and optical axis position of the camera,and converted into the spindle in the world coordinate system Pan rotation information.In this paper,through the spindle position experiment and analysis of key point detection and target detection,finally combining the advantages of multiple detection methods,based on the detection method of YOLO,a network model(BKNet)for simultaneous target detection and key point detection is proposed to Acquire the twodimensional spatial information of the spindle in the color image.At the same time,in order to speed up the running speed of the model,this paper uses the attention module(SE Module)based on the feature map channel to select important channels and accelerate the model by reducing the channel.Then continue to use lightweight model structure optimization and network pruning methods to improve the speed of model detection.At the same time,this paper proposes to detect the spindle detection strategy selected multiple times to improve the efficiency of spindle position detection.According to the two-dimensional spatial information of the spindle obtained from the detection of the color image by the BKNet model,as many subsequent spindles as possible are selected to reduce the number of model runs and save time overhead when detecting the same amount of spindles.In this paper,the depth map obstruction detection method and the frame difference method are used to judge the influence on the positions of other spindles after removing the candidate spindles to ensure the accuracy of the detection results.